Abstract: Forced oscillation has the characteristics of strong randomness, large amplitude, fast start-up and wide propagation range. In order to prevent power system splitting and blackout caused by forced oscillation, it is very important and urgent to locate the oscillation source quickly and accurately. Based on the wide area measurement data, a data-driven forced oscillation source location method is proposed in this paper. Firstly, the isolation forest algorithm is used to clean the original data, and then the original problem is transformed into a robust principal component analysis problem by using the sparsity of the forced oscillation data and the low rank of the original system data, which is solved by the augmented Lagrange multiplier method. Finally, the effectiveness of this method is verified on WECC 240 node system.